Knowledge acquisition: automated approaches

1
Knowledge acquisition automated approaches

• scenario  have many examples of the decision
  i.e. inputs and outputs but no access to
  expertise or a decision maker.
• or the decision maker expert is inarticulate
• Can induce a decision tree from the data. A set
  of production rules can be generated from the
  decision tree.
• Induction means reasoning from the specific (the
  examples) to the general (the rule set).
• Several algorithms have been developed - this is
  normally referred to as machine learning

2
Knowledge acquisition automated approaches

• A manual example A loan application advisor
• the following table contains examples of the
  decision, can we induce any knowledge from it?

3
Knowledge acquisition automated approaches

• We could infer the following 3 rules
• 1 IF income is 70 000 or more THEN approve
  the loan
• explains' Ms Green and Ms Brown
• 2 IF income is 30 000 or more AND age is
  at least 42 AND there are no dependents AND
  assets are at least 250 000 THEN approve the
  loan
• explains' Ms Rich, Mr Black, Mr Smith
• 3 IF income is between 30 000 and 50 000
  
• AND assets are at least 100 000 THEN
  approve the loan
• explains' Mr White and Mr Smith

4
Knowledge acquisition automated approaches

• These are examples of production' rules.
• They express knowledge'
• They can only tell us whether we should approve
  the loan.
• When all clauses on the left hand side of a rule
  are true then the rule fires
• and provides a decision.
• The order in which the rules are processed' can
  vary.

5
Knowledge acquisition automated approaches

• This is different than writing an selection block
  using the IF statement
• i.e
• IF income is 70 000 or more
• then approve the loan
• ELSE
• IF income is 30 000 or more
• AND age is at least 42
• AND there are no dependents
• AND assets are at least 250 000
• then approve the loan
• ELSE
• IF income is between 30 000
  and 50 000
• AND assets are at least 100
  000
• then approve the loan
• ELSE
• reject loan
• ENDIF
• ENDIF
• ENDIF

6
Knowledge acquisition automated approaches

• What about the following examples?
• 1 Mr Blue who earns 45 000 p.a. has assets
  worth 110, 000,
• is 30 years of age and has no dependents
• applies for a loan and is rejected.
• problem the rule base approves the loan
  rule 3 fires
• could add an extra clause to rule 3 i.e.
  AND has at least one dependent
• 2 Mrs Mayberry who earns 70 000, has no
  assets,
• is 35 years old and has one dependent
• applies for a loan and is also rejected
• a more serious problem we have conflict
  in our examples

7
Knowledge acquisition

• Multiple Experts
• may have more than one expert to deal with
• How to you integrate their opinion?
• How do you deal with conflict or
  disagreement?
• How do you handle multiple lines of
  reasoning?

8
Knowledge representation analysis representation
Interviews Protocol analysis Repertory grid
Induction etc
9
Analysis representation

• Problem
• What is an effective way of documenting what we
  are finding out about
• the problem domain and the decision making?
• Hopefully the transition from documentation to
  knowledge representation (and consequently
  implementation) is reasonably straight forward.
• Some knowledge acquisition techniques
• particularly the semi-automated and automated
  ones
• generate their own documentation techniques

10
Analysis representation

• For example
• An automated technique such as induction can
  generate a decision tree, which can be expressed
  as a selection block (i.e. a series of nested IF
  statements) which in turn can be written as a set
  of production rules.
• A semi-automated technique such repertory grid
  analysis can generate schematics such as rating
  grids and distance matrices.
• What about results from the various manual
  acquisition techniques?

11
Analysis representation

• For declarative/'factual' knowledge
• You could consider the use of
• a glossary meaning of technical terms
  (or concepts)
• classification trees group concepts
  together
• according to common properties.
• and /or a conceptual graph/ semantic
  network approach
• documents the relationships between
  concepts in the problem domain.
• different relationship types can be
  documented
• A has an attribute B
• A has an association with B
• A is a component of B
• A takes place in B etc..

12
Analysis representation

• Techniques such as conceptual graphs, semantic
  networks, influence diagrams etc
• can be formal or informal
• formality implies that there is a syntax (or
  grammar')
• and a set of rules governing how the graph (or
  network) can be developed.
• This is useful because it implies that the
  representation scheme may be implementable (in a
  computing sense)
• informality (or perhaps semi-informality)
  implies only general guidelines with some rules.
• Generally such schematics are not implementable.
  

13
Analysis representation

• An example of an informal conceptual
  graph/semantic network for a hospital

key
concept
hospital
is-part-of
the health system
relationship
contains
contains
direction
surgeon
ward
operating theatre
uses
attr
attr
operates on
located in
patient
bed
ward
number of beds